Add dataloaders for handling dataset for torch modules

[ankitpatnala]

There are two dataloading type defined.
-> Use map based dataset (seemed effective for parallelizing)
-> Use iterable dataset (seemed effective with single node) (Further can be looked upon to update to multi-node multi-gpu setting)

[seblehner]

@ankitpatnala Is this issue still open/relevant, or can we close?